A microgrid robust economic dispatch method based on a set of probability distributions

Abstract

This invention discloses a robust economic dispatch method for microgrids based on probability distribution sets. Addressing the challenge of balancing dispatch economy and safety due to the uncertainty of high-proportion renewable energy sources, it extracts statistical characteristics of renewable energy and loads, generates uncertainty scenarios, and constructs an initial fuzzy set. Through extreme scenario safety testing, it eliminates risk probability distributions that inevitably lead to operational overruns, forming a robust probability distribution feasible region that eliminates safety risks at the source. A two-stage sub-Bruker optimization model is established to solve for the optimal dispatch strategy under the worst-case probability distribution within the feasible region, employing an iterative decomposition algorithm for efficient solution. Before outputting dispatch instructions, a forward-looking extrapolation verification and data traceability repair mechanism is introduced, forming an intelligent closed loop of optimization verification and traceability learning, driving adaptive adjustment of key parameters. This invention balances robust safety and operational economy, improving the long-term adaptability of microgrid dispatch under complex uncertainty environments.

Description

Technical Field

[0001] This invention belongs to the field of robust economic dispatch technology for microgrids, and in particular relates to a robust economic dispatch method for microgrids based on probability distribution sets. Background Technology

[0002] As the global energy structure transitions towards green and low-carbon practices, the penetration rate of renewable energy sources, such as photovoltaics and wind power, in microgrids continues to increase. However, both renewable energy output and load demand exhibit significant randomness and volatility, posing serious challenges to the safe, economical, and reliable operation of microgrids. Economic dispatch is the core of microgrid operation, aiming to optimize the output of various controllable resources and minimize total operating costs while meeting system security constraints. To address uncertainty, existing dispatch methods mainly include stochastic optimization, robust optimization, and the recently developed sub-Bruker optimization.

[0003] Stochastic optimization methods rely on precise probability distribution models of uncertain parameters. When the accurate distribution is known, these methods can achieve economically efficient scheduling schemes. However, in practice, precise probability distributions are difficult to obtain, and distributions estimated based on limited historical data or predictive information often contain biases. When there is a significant difference between the actual distribution and the preset distribution, stochastic optimization schemes may lead to serious operational risks due to overly optimistic estimations of uncertainty, such as equipment overload and voltage exceeding limits. Robust optimization methods, on the other hand, take a completely conservative approach, assuming that the uncertain parameters fluctuate arbitrarily within a preset, defined range, and optimizing system performance under worst-case scenarios. While this method provides strict safety guarantees, its results are often overly conservative. This is because it sacrifices the economic potential under normal conditions to defend against theoretically possible but practically low-probability worst-case scenarios, resulting in high scheduling costs and poor economic efficiency.

[0004] Partially Brutal optimization (BBO) is a method that lies between stochastic optimization and robust optimization. It does not assume that uncertainty follows a single exact distribution, but rather that the true distribution belongs to a feasible region of probability distribution centered on a reference distribution. This method aims to optimize the expected performance under the worst-case distribution, thus balancing economy and robustness to a certain extent. However, existing microgrid dispatching research based on BBO often relies on statistical moments or probability divergence to construct the feasible region of probability distribution. While the fuzzy sets constructed by these methods include statistically "possible" distributions, they may still contain a large number of probability distributions that, while satisfying statistical constraints, would inevitably lead to system overruns in actual physical operations. Using such feasible regions for optimization, the resulting "robust" strategies still carry security risks. Existing methods are typically an "open-loop," one-time optimization process, directly outputting dispatch commands after model solving, lacking final feasibility verification of the optimization results in real-world or high-fidelity simulation environments. When strategy failure occurs due to data quality issues, model bias, or sudden anomalies, the system cannot automatically diagnose the root cause, repair data defects, or adjust model parameters, making it difficult to adapt to complex and ever-changing real-world operating environments, potentially leading to long-term performance degradation. Therefore, a robust and economical dispatch method for microgrids based on probability distribution sets needs to be proposed to solve the above problems. Summary of the Invention

[0005] The technical problem this invention aims to solve is to provide a robust and economical microgrid dispatch method based on probability distribution sets. This method innovatively integrates safety-oriented robust optimization with data-driven closed-loop learning. By constructing a robust probability distribution feasible region that has undergone safety testing and is compacted, high-risk probability distributions are eliminated at the source. A two-stage sub-Bruker robust optimization model is established to solve the optimal dispatch strategy for the worst probability distribution. Furthermore, an intelligent closed loop of optimization-verification-source tracing-repair-learning is designed. Before outputting instructions, high-fidelity simulation is used for forward-looking deduction and verification. Data tracing and adaptive learning of model parameters are performed on discovered problems. This provides systematic technical support for solving the key technical problem of balancing economy, robustness, and long-term adaptability in microgrid dispatch under high-proportion renewable energy access.

[0006] To achieve the above technical solution, the technical solution adopted by the present invention is as follows: A robust and economical dispatch method for microgrids based on probability distribution sets includes the following steps: S1. Acquire and standardize historical and real-time photovoltaic and wind power output and load data, and quantify uncertainty characteristics by analyzing the data fluctuation patterns and statistical properties; S2. Establish an optimization objective function with the goal of minimizing the total operating cost of the microgrid system; S3. Construct a feasible domain of a safe and robust probability distribution for subsequent robust optimization. First, generate uncertain scenarios and construct an initial fuzzy set describing the possible fluctuation range of the probability distribution. Through safety testing based on extreme scenarios, remove all risk probability distributions that will inevitably lead to the system running beyond its limits from this set, thereby forming a robust probability distribution feasible domain that eliminates security risks from the source and is compacted. S4. After generating a large number of initial uncertain scenarios through step S3, the numerous initial scenarios are merged and filtered according to their probability characteristics through scenario reduction technology, and finally reduced to a limited number of representative key scenarios, and their corresponding adjusted probabilities are calculated. S5. Establish a two-stage sub-Bruker optimization model and solve the optimal scheduling strategy for dealing with the worst probability distribution of uncertainty within the feasible region of the robust probability distribution. S6. Solve the two-stage sub-Brussels bar optimization model established in step S5 using an iterative decomposition algorithm; S7. Before outputting the final scheduling instruction, perform simulation processing to verify the feasibility of the strategy and generate an anomaly detection report. S8. Using the anomaly detection report generated in step S7, trace back and locate the source data that caused the scheduling strategy problem. S9. For the problematic data identified and isolated in step S8, initiate the automatic repair and enhancement process; S10. Use the closed-loop feedback information formed in steps S7 to S9 as a continuous monitoring signal to drive the adaptive learning of key parameters in the scheduling method. S11. After performing the solution process of step S6 and comprehensively considering the closed-loop feedback and dynamic adjustment results of steps S7 to S10, the first-stage day-ahead scheduling plan and the second-stage real-time adjustment strategy that minimize the total cost of the system under the worst probability distribution of uncertainty and pass the look-ahead verification are obtained, namely the optimal scheduling strategy. S12, Output dispatch instructions for each controllable power generation unit, energy storage system and demand response resource.

[0007] Preferably, step S1 specifically includes the following steps: A1. Obtain historical and real-time photovoltaic power output data, wind power output data, and load demand data to form the raw input dataset; A2. Perform preprocessing on the original input dataset. Preprocessing specifically includes: A21. Identify and remove obvious errors and outliers from the data; A22. For missing time periods in the data, use interpolation of nearby data or filling methods based on the historical average of similar days to complete the data; A23. Scale all data to the [0,1] interval according to their respective maximum and minimum value ranges to form a standardized input data sequence; A3. Calculate the statistical characteristics of uncertainty parameters using standardized input data sequences: A31. Calculate the historical maximum and minimum value ranges of photovoltaic power output, wind power output, and load demand during typical hourly periods; A32. Calculate the Pearson correlation coefficient matrix between photovoltaic and wind power output, and between renewable energy output and load demand; A33. For the prediction errors and load fluctuations of photovoltaic and wind power respectively, the probability distributions are fitted using historical data. The distributions include, but are not limited to, normal distribution, Beta distribution, and Gaussian mixture distribution, and the probability characteristics are characterized by the distribution parameters. A4. Directly output the processed standardized data sequence, as well as the set of uncertain statistical features of photovoltaic, wind power output and load demand, including their fluctuation range, correlation coefficient matrix and fitted probability distribution parameters, as input for generating uncertain scenarios.

[0008] Preferably, the total operating cost in step S2 specifically includes the cost of purchasing and selling electricity, the cost of energy storage equipment losses, the generation cost of controllable distributed power sources, and the regulation cost generated by implementing demand response; the solution of the objective function needs to satisfy the system power balance constraints, the physical constraints of operation of each power generation unit and energy storage equipment, and the safety constraints to ensure system operation.

[0009] Preferably, step S3 specifically includes the following steps: S31. Using the probability distribution features extracted in step S1, generate typical scenarios that characterize the possible future states of the uncertainty parameters, and assign an initial probability estimate to each scenario. S32. Measure the deviation between the overall probability distribution of the generated scene set and a preset reference distribution, and construct a fuzzy set that allows the true probability to fluctuate around the reference distribution. The fuzzy set defines the feasible region of the initial probability distribution formed by all possible probability distributions. S33. Using historical extreme scenarios or a pre-set set of security test scenarios, pre-screen the feasible region of the initial probability distribution and remove those probability distributions that, although they meet the definition of fuzzy sets, would cause any feasible scheduling strategy to violate key security constraints. S34. Through the pre-screening step, the initial fuzzy probability feasible region is compressed into a robust probability distribution feasible region that ensures scheduling safety, providing an uncertainty description framework.

[0010] Preferably, step S5 specifically includes the following steps: S51. To make decisions in response to uncertainty, the scheduling decision variables are divided into two stages; S52. The first stage consists of day-ahead decision variables, which are determined before the uncertainty is revealed; the second stage consists of real-time adjustment variables, which are adjusted after the uncertainty is realized, and a two-stage sub-Bruker optimization model is established. S53. The objective of the sub-Bruker optimization model is to find the optimal scheduling strategy that can cope with the worst probability distribution situation within the feasible region of the robust probability distribution defined in step S3. S54. The mathematical form of the sub-Bruker optimization model is as follows: determine the first-stage decision variables and their costs; for the first-stage decision and any possible uncertainty scenario, optimize the second-stage decision variables to minimize the operating cost under that scenario; the optimization objective is to minimize the sum of the first-stage cost and the expected second-stage cost under the worst probability distribution in the feasible region of the robust probability distribution.

[0011] Preferably, step S6 specifically includes the following steps: S61. Solving the main problem: Using the key scenario and adjusted probability of the current step S4, solve the deterministic two-stage optimization problem, obtain the candidate solution for the first-stage decision, and calculate the lower bound of the objective function. S62. Subproblem solving: For the first-stage candidate solutions given by the main problem, within the feasible region of the robust probability distribution defined in step S3, find the worst-case probability distribution that maximizes the total expected cost, and calculate the corresponding expected cost value to form the upper bound of the objective function. S63. Iteration and Convergence: Compare the upper and lower bounds obtained from the current calculation. If the difference is less than the preset convergence tolerance, the algorithm terminates and the current solution is the optimal scheduling strategy. Otherwise, add the new worst-case scenario or the corresponding probability distribution information identified in the subproblem to the scenario set of the main problem, solve the main problem again, and repeat the above process until convergence to seek the global optimal solution. S64. During the iterative solution process, an online evaluation mechanism is introduced to assess the coupling relationship between the day-ahead decision in the first stage and the real-time adjustment decision variables in the second stage. S65. When the evaluation shows that strong coupling between the two leads to difficulty in solving the problem, the expression of the optimization model is dynamically adjusted, and a decoupling and simplification strategy is implemented between stages in some time periods or for some devices to improve the solution efficiency; when the evaluation shows that strong coupling is crucial to ensuring the robustness of the strategy, the coupling relationship is maintained or strengthened.

[0012] Preferably, step S7 specifically includes the following steps: S71. Before outputting the final scheduling instruction in step S6, the generated candidate optimal scheduling strategy is prospectively simulated in a high-fidelity simulation environment built based on the current data. S72. The forward-looking simulation simultaneously monitors and records all potential problems that may occur during the strategy execution process, such as constraint overruns, equipment overloads, and voltage / frequency anomalies, as well as key uncertainty scenario segments and input data characteristics that lead to these problems. S73. Finally, generate an anomaly detection report.

[0013] Preferably, the source data includes outliers in the original input data, data points that may be incorrectly corrected in the preprocessing step S1, and key uncertainty scenarios that lead to risks; all identified problem data and context information are automatically backed up and stored in the problem data and scenario library, and are physically or logically isolated from the normal data sources used in the main process.

[0014] Preferably, step S9 specifically includes the following steps: S91. Initiate an automatic repair and enhancement process for problematic data. For data with obvious errors or missing information, use intelligent interpolation and correction algorithms for repair. S92. For uncertain scenario data that repeatedly causes policy risks but is not obvious errors, special labeling and enhancement are performed by increasing the corresponding penalty weight in the objective function, or by generating adversarial samples based on it. S93. These enhanced data are used as a special stress test scenario and fed back to the probability distribution feasible domain construction in step S3 or the key scenario set in step S4.

[0015] Preferably, step S10 specifically includes the following steps: S101. The closed-loop feedback information formed in steps S7 to S9 is used as a continuous monitoring signal to drive the adaptive learning of key parameters in the scheduling method. S102. Establish a feedback learning mechanism to dynamically adjust the data cleaning threshold in step S1, the confidence radius of the feasible region of the probability distribution in step S3, the strictness of the safety constraints in step S3, and the tolerance core parameters of the model convergence judgment in step S6, based on the frequency, type, and repair effect of anomalies in the historical cycle. S103, enabling the entire scheduling system to continuously learn from historical problems and optimize itself.

[0016] The beneficial effects of this invention are as follows: 1. When constructing the uncertainty model, this method not only considers the statistical characteristics of the probability distribution, but also introduces a safety testing mechanism. It actively removes risk distributions that will inevitably lead to the system running out of limits from the initial probability distribution feasible region, and constructs a compressed robust probability distribution feasible region. This makes subsequent optimization no longer target all possible probability distributions, but specifically targets an acceptable and safe set of uncertainties for robust optimization. The final scheduling strategy is mathematically guaranteed to meet the system safety constraints under all possible probability distributions within the compressed feasible region. Thus, in complex and ever-changing uncertain environments, it fundamentally enhances the robustness and safety assurance level of scheduling decisions. 2. Through scenario reduction in step S4 and intelligent iterative solution in step S6, this method significantly improves computational efficiency while ensuring optimization quality. The compressed feasible region narrows the decision search space, making the solution more focused. Furthermore, the two-stage sub-Bruker optimization model and its solution algorithm can optimize the expected total cost while considering the worst-case probability distribution. The online evaluation and dynamic decoupling mechanism of coupling relationships introduced in step S6 further achieves an intelligent balance between solution accuracy and efficiency. 3. By forward-looking simulation to verify the feasibility of the strategy, problems are traced back to the source data and repaired and enhanced. Finally, the closed-loop feedback information is used to drive the adaptive learning of key model parameters, so that the entire scheduling system is no longer a static model, but an intelligent agent that can continuously learn from historical operational anomalies and dynamically adjust its own parameters. It has the ability to continuously evolve and self-optimize from actual operating experience, and can adapt to changes in data characteristics and external environment evolution in the long term, so as to achieve continuous performance improvement and maintenance. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the main process of the present invention; Figure 2 This is a schematic diagram of the initial fuzzy process of the present invention; Figure 3 This is a schematic diagram comparing the voltage safety robustness distribution under different scheduling methods in the embodiments of the present invention; Figure 4 This is a schematic diagram showing a comprehensive comparison of the multi-dimensional performance indicators of the four scheduling methods in this embodiment of the invention. Detailed Implementation

[0018] Example 1: like Figure 1 and Figure 2 As shown, a robust and economical dispatch method for microgrids based on probability distribution sets includes the following steps: S1. Acquire and standardize historical and real-time photovoltaic and wind power output and load data, and quantify uncertainty characteristics by analyzing the data fluctuation patterns and statistical properties; S2. Establish an optimization objective function with the goal of minimizing the total operating cost of the microgrid system; S3. Construct a feasible domain of a safe and robust probability distribution for subsequent robust optimization. First, generate uncertain scenarios and construct an initial fuzzy set describing the possible fluctuation range of the probability distribution. Through safety testing based on extreme scenarios, remove all risk probability distributions that will inevitably lead to the system running beyond its limits from this set, thereby forming a robust probability distribution feasible domain that eliminates security risks from the source and is compacted. S4. After generating a large number of initial uncertain scenarios through step S3, the numerous initial scenarios are merged and filtered according to their probability characteristics through scenario reduction technology, and finally reduced to a limited number of representative key scenarios, and their corresponding adjusted probabilities are calculated. S5. Establish a two-stage sub-Bruker optimization model and solve the optimal scheduling strategy for dealing with the worst probability distribution of uncertainty within the feasible region of the robust probability distribution. S6. Solve the two-stage sub-Brussels bar optimization model established in step S5 using an iterative decomposition algorithm; S7. Before outputting the final scheduling instruction, perform simulation processing to verify the feasibility of the strategy and generate an anomaly detection report. S8. Using the anomaly detection report generated in step S7, trace back and locate the source data that caused the scheduling strategy problem. S9. For the problematic data identified and isolated in step S8, initiate the automatic repair and enhancement process; S10. Use the closed-loop feedback information formed in steps S7 to S9 as a continuous monitoring signal to drive the adaptive learning of key parameters in the scheduling method. S11. After performing the solution process of step S6 and comprehensively considering the closed-loop feedback and dynamic adjustment results of steps S7 to S10, the first-stage daytime scheduling plan and the second-stage real-time adjustment strategy that minimize the total cost of the system under the worst probability distribution of uncertainty and pass the look-ahead verification are obtained, namely the optimal scheduling strategy. S12, Output dispatch instructions for each controllable power generation unit, energy storage system and demand response resource.

[0019] Through the scenario reduction technique in step S4 and the iterative decomposition algorithm in step S6, this method achieves an efficient solution to complex problems. S4 condenses massive scenarios into representative key scenarios, significantly reducing the model size. S6 employs a master-subproblem iterative framework, continuously identifying and adding the worst-case scenarios, ensuring that the solution approaches the global optimum within the compressed feasible region. In particular, the online evaluation and dynamic adjustment mechanism for coupling relationships introduced in steps S64 and S65 intelligently balances solution accuracy and computational efficiency based on the problem structure, making this complex sub-Bruker optimization model engineering-practical and capable of outputting a high-quality scheduling plan that balances economy, robustness, and safety.

[0020] Step S1 specifically includes the following steps: A1. Obtain historical and real-time photovoltaic power output data, wind power output data, and load demand data to form the raw input dataset; A2. Perform preprocessing on the original input dataset. Preprocessing specifically includes: A21. Identify and remove obvious errors and outliers from the data; A22. For missing time periods in the data, use interpolation of nearby data or filling methods based on the historical average of similar days to complete the data; A23. Scale all data to the [0,1] interval according to their respective maximum and minimum value ranges to form a standardized input data sequence; A3. Calculate the statistical characteristics of uncertainty parameters using standardized input data sequences: A31. Calculate the historical maximum and minimum value ranges of photovoltaic power output, wind power output, and load demand during typical hourly periods; A32. Calculate the Pearson correlation coefficient matrix between photovoltaic and wind power output, and between renewable energy output and load demand; A33. For the prediction errors and load fluctuations of photovoltaic and wind power respectively, use historical data to fit the probability distribution. The distribution includes, but is not limited to, normal distribution, Beta distribution, and Gaussian mixture distribution. The probability characteristics are characterized by the distribution parameters. A4. Directly output the processed standardized data sequence, as well as the uncertainty statistical feature set of photovoltaic, wind power output and load demand, including its fluctuation range, correlation coefficient matrix and fitted probability distribution parameters, as the input basis for generating uncertainty scenarios. The data processing workflow was standardized, and the quality and comparability of the input data were ensured through data cleaning, missing value imputation, and standardization. Step A3 systematically extracted multi-dimensional statistical features such as fluctuation range, correlation coefficient, and probability distribution parameters, providing accurate and reliable quantitative input for generating uncertainty scenarios that conform to real statistical laws and constructing feasible domains of probability distributions in subsequent steps S3 and S4. This is the reliable data foundation for the effective implementation of the entire method.

[0021] Specifically, the total operating cost in step S2 includes the cost of purchasing and selling electricity, the cost of energy storage equipment losses, the generation cost of controllable distributed power sources, and the regulation cost generated by implementing demand response. The solution of the objective function must satisfy the system power balance constraints, the physical constraints of operation of each power generation unit and energy storage equipment, and the safety constraints to ensure the operation of the system. The specific components of the total cost in the optimization objective function are clearly defined, including electricity purchase and sale costs, equipment losses, generation costs, and demand response control costs, thus specifying the economic objective. Simultaneously, the system power balance, equipment physical constraints, and safety constraints that must be satisfied in the solution are stipulated, establishing a complete constraint system. This ensures that the optimization model closely aligns with the economic and physical laws of actual microgrid operation, making the obtained scheduling strategy not only mathematically optimal but also practically feasible in engineering, guaranteeing the practical application value of the solution.

[0022] Step S3 specifically includes the following steps: S31. Using the probability distribution features extracted in step S1, generate typical scenarios that characterize the possible future states of the uncertainty parameters, and assign an initial probability estimate to each scenario. S32. Measure the deviation between the overall probability distribution of the generated scene set and a preset reference distribution, and construct a fuzzy set that allows the true probability to fluctuate around the reference distribution. The fuzzy set defines the feasible region of the initial probability distribution formed by all possible probability distributions. S33. Using historical extreme scenarios or a pre-set set of security test scenarios, pre-screen the feasible region of the initial probability distribution and remove those probability distributions that, although they meet the definition of fuzzy sets, would cause any feasible scheduling strategy to violate key security constraints. S34. Through the pre-screening step, the initial fuzzy probability feasible region is compressed into a robust probability distribution feasible region that ensures scheduling safety, providing an uncertainty description framework. The key steps for constructing a robust probability distribution feasible region are refined. First, typical scenarios and initial fuzzy sets are generated based on statistical features. Then, a security test pre-screening step is innovatively introduced to eliminate distributions that violate security constraints using extreme scenarios. Finally, a compressed secure feasible region is obtained. This process directly embeds physical security requirements into the mathematical framework of uncertainty description. It is a concrete implementation of claim 1 to eliminate security risks from the source and is the core technical link to improve the security robustness of the entire method.

[0023] Step S5 specifically includes the following steps: S51. To make decisions in response to uncertainty, the scheduling decision variables are divided into two stages; S52. The first stage consists of day-ahead decision variables, which are determined before the uncertainty is revealed; the second stage consists of real-time adjustment variables, which are adjusted after the uncertainty is realized, and a two-stage sub-Bruker optimization model is established. S53. The goal of the split-bar optimization model is to find the optimal scheduling strategy that can cope with the worst probability distribution situation within the feasible region of the robust probability distribution defined in step S3. S54. The mathematical form of the sub-Bruker optimization model is as follows: Determine the decision variables and their costs in the first stage; for the first stage decision and any possible uncertainty scenario, optimize the decision variables in the second stage to minimize the operating cost in that scenario; the optimization objective is to minimize the sum of the first stage cost and the expected second stage cost under the worst probability distribution in the feasible region of the robust probability distribution. The two-stage structure of the sub-Bruker optimization model is clearly defined: day-ahead decision and real-time adjustment. This structure accurately describes the timing decision-making logic of advance planning and real-time response in microgrid dispatch. The model objective is to optimize the expected total cost under the worst-case distribution within the safe and feasible domain. It considers multiple possibilities of uncertainty and defends against the most unfavorable situation. Mathematically, it rigorously balances economy and robustness, providing an accurate model framework for solving high-quality dispatch strategies.

[0024] Step S6 specifically includes the following steps: S61. Solving the main problem: Using the key scenario and adjusted probability of the current step S4, solve the deterministic two-stage optimization problem, obtain the candidate solution for the first-stage decision, and calculate the lower bound of the objective function. S62. Subproblem solving: For the first-stage candidate solutions given by the main problem, within the feasible region of the robust probability distribution defined in step S3, find the worst-case probability distribution that maximizes the total expected cost, and calculate the corresponding expected cost value to form the upper bound of the objective function. S63. Iteration and Convergence: Compare the upper and lower bounds obtained from the current calculation. If the difference is less than the preset convergence tolerance, the algorithm terminates and the current solution is the optimal scheduling strategy. Otherwise, add the new worst-case scenario or the corresponding probability distribution information identified in the subproblem to the scenario set of the main problem, solve the main problem again, and repeat the above process until convergence to seek the global optimal solution. S64. During the iterative solution process, an online evaluation mechanism is introduced to assess the coupling relationship between the day-ahead decision in the first stage and the real-time adjustment decision variables in the second stage. S65. When the evaluation shows that strong coupling between the two leads to difficulty in solving the problem, the expression of the optimization model is dynamically adjusted, and a decoupling and simplification strategy is implemented between stages in certain time periods or for certain devices to improve the solution efficiency; when the evaluation shows that strong coupling is crucial to ensuring the robustness of the strategy, the coupling relationship is maintained or strengthened. By iterating through the main problem and subproblems, the algorithm continuously approaches the optimal solution and uses upper and lower bounds to determine convergence, ensuring the accuracy of the solution. The online evaluation and dynamic adjustment mechanism of coupling relationships introduced in steps S64 and S65 is an intelligent feature. It can dynamically decide whether to decouple and simplify the model to improve efficiency or maintain coupling to ensure accuracy based on the current solution state. This enables the algorithm to adaptively handle problems of different complexities and obtain high-quality solutions within an acceptable time, which is the key to the practical application of complex models.

[0025] Step S7 specifically includes the following steps: S71. Before outputting the final scheduling instruction in step S6, the generated candidate optimal scheduling strategy is prospectively simulated in a high-fidelity simulation environment built based on the current data. S72. The forward-looking simulation simultaneously monitors and records all potential problems that may occur during the strategy execution process, such as constraint overruns, equipment overloads, and voltage / frequency anomalies, as well as key uncertainty scenario segments and input data characteristics that lead to these problems. S73. Finally, generate the anomaly detection report; By placing the optimized strategy in a high-fidelity simulation environment for dynamic execution, runtime issues that may be overlooked in static optimization, such as dynamic limit violations and equipment overload, can be discovered and recorded in advance. This is equivalent to adding a reliable safety valve and testing step before actual execution, which can effectively intercept scheduling instructions with potential risks and greatly improve the reliability and security of scheduling scheme delivery.

[0026] The source data includes outliers in the original input data, data points that may be incorrectly corrected in step S1 preprocessing, and key uncertainty scenarios that lead to risks. All identified problem data and context information will be automatically backed up and stored in the problem data and scenario library, and physically or logically isolated from the normal data sources used by the main process. The scope of problematic data was clearly defined, including original anomalies, erroneous correction points, risk scenarios and their handling methods. These were backed up and stored in an independent database to achieve physical or logical isolation. This not only protected the cleanliness of the main process data source and prevented contamination, but also established a dedicated problematic case database, providing clear and safe objects and targets for subsequent data repair, root cause analysis and system learning. This is an important data foundation for achieving intelligent closed-loop management.

[0027] Step S9 specifically includes the following steps: S91. Initiate an automatic repair and enhancement process for problematic data. For data with obvious errors or missing information, use intelligent interpolation and correction algorithms for repair. S92. For uncertain scenario data that repeatedly causes policy risks but is not obvious errors, special labeling and enhancement are performed by increasing the corresponding penalty weight in the objective function, or by generating adversarial samples based on it. S93. Treat these enhanced data as a special stress test scenario and feed them back to the probability distribution feasible region construction in step S3 or the key scenario set in step S4. For obvious errors, they are fixed; for data or scenarios that are not obvious but pose risks, they are specially marked and enhanced, such as by increasing penalty weights or generating adversarial samples. This process not only fixes data defects, but also transforms known risks into stress test scenarios that enhance the robustness of the model. Feeding these enhanced data back to the feasible domain to construct S3 or scenario set S4 enables the optimization model to actively remember and defend against similar risks, demonstrating the system's ability to learn and evolve from errors.

[0028] Specifically, step S10 includes the following steps: S101. The closed-loop feedback information formed in steps S7 to S9 is used as a continuous monitoring signal to drive the adaptive learning of key parameters in the scheduling method. S102. Establish a feedback learning mechanism to dynamically adjust the data cleaning threshold in step S1, the confidence radius of the feasible region of the probability distribution in step S3, the strictness of the safety constraints in step S3, and the tolerance core parameters of the model convergence judgment in step S6, based on the frequency, type, and repair effect of anomalies in the historical cycle. S103. Enable the entire scheduling system to continuously learn from historical problems and optimize itself. By using information such as the frequency, type, and repair effect of anomalies generated in the closed loop as monitoring signals, multiple core model parameters are dynamically adjusted, such as data cleaning threshold, feasible region confidence radius, safety constraint strictness, and convergence tolerance. This transforms the entire scheduling system from a static model with fixed parameters into an intelligent agent capable of automatically optimizing based on historical performance. The system can adapt to changes in data distribution and environmental evolution, achieving continuous self-optimization of performance and possessing the core competitiveness to maintain high efficiency and reliable operation in the long term.

[0029] Example 2: This embodiment selects a photovoltaic-storage microgrid in an industrial park in southern China as the test object. This park contains high-energy-consuming production lines and office loads, requiring extremely high power supply reliability. The system includes an 800kWp photovoltaic array, a 1.5MW / 3MWh lithium iron phosphate energy storage system, and a 500kW micro gas turbine (MT), connected to the main grid via a 10kV bus and a 1MW tie line. Some non-critical loads in the park, including air conditioning and some lighting, have signed demand response (DR) agreements, with a maximum load reduction of 200kW. The dispatch cycle is for the next 24 hours, with a time resolution of 1 hour.

[0030] System / method settings parameters: Time-of-use electricity pricing: During peak hours (10:00-13:00, 18:00-21:00), the purchase price is 1.2 yuan / kWh; during normal hours, it is 0.8 yuan / kWh; and during off-peak hours (23:00-07:00), it is 0.4 yuan / kWh.

[0031] Equipment cost parameters: Gas turbine power generation cost coefficients a=0.0012 yuan / kW², b=0.65 yuan / kW, c=20 yuan; Energy storage charging and discharging loss cost: 0.05 yuan / kWh; Demand response compensation price: 1.5 yuan / kWh.

[0032] Algorithm parameters: The confidence interval in step S1 is set to 95%; the Wasserstein fuzzy set radius in step S3 is... =0.1; the convergence tolerance in step S6 is set to 0.5%.

[0033] To verify the advancement of the method of this invention, the following three prior art technologies are set as comparison benchmarks: Traditional stochastic optimization (SO): It assumes that the uncertainty parameters strictly follow a normal distribution, does not consider the fuzziness of the distribution, and aims to minimize the expected cost.

[0034] Traditional Robust Optimization (RO): Based on a box-type uncertainty set (PV / load fluctuation range is ±20% of the predicted value), optimize the cost under the worst single scenario.

[0035] Standardized Divided Bar Optimization (DRO): Employs a fuzzy set construction method based on moment information, but does not include the S3 safe compaction step and the S7-S10 closed-loop feedback learning mechanism of this invention.

[0036] Test and comparison running conditions: To ensure a fair comparison, all methods were run on the same test dataset for 30 consecutive running days (covering typical days of sunny, cloudy, rainy, and load abrupt changes). The test platform was MATLAB R2022a, using the CPLEX 12.10 solver. The test data included actual prediction error sequences and extreme weather disturbances.

[0037] The specific operation process is as follows: First, data preprocessing and uncertainty quantification are performed. Historical and predicted photovoltaic (PV) output, wind power output, and load data are obtained from the energy management system and weather stations to form the original input dataset. For historical data, box plots are used to identify and remove outliers such as non-zero output at night. For missing data, linear interpolation of data from the preceding and following two hours is used to complete the data. All time-series data are divided by their historical maximum observed value and standardized to the zero-to-one range to form a standardized input data sequence. Based on this standardized sequence, the statistical characteristics of uncertainty parameters are calculated, including the fluctuation range of PV, wind power, and load over 24 typical hours each day, such as the midday PV output range. The range is 0.7 to 1.0. The Pearson correlation coefficient matrices between photovoltaic (PV) and wind power output, and between renewable energy output and load demand, are calculated. PV and wind power show a weak negative correlation (coefficient -0.2), while PV and daytime load show a positive correlation (coefficient 0.6). Probability distribution fitting is performed on the prediction error sequences of PV and wind power, and the load fluctuation sequence. It is found that the prediction errors of wind and solar power follow a mixed Gaussian distribution with a mean of zero and different variances, while the load fluctuation approximately follows a normal distribution. The parameters of these distributions are recorded. Finally, the processed data sequence and a set of uncertainty statistical features including fluctuation ranges, correlation coefficient matrices, and distribution parameters are output as input for generating uncertainty scenarios. Construct a robust probability distribution feasible region and key scenarios; using the previously extracted Gaussian mixture distribution parameters and correlation matrix, adopt Latin hypercube sampling combined with Cholesky decomposition to generate 10,000 uncertain scenarios representing the joint fluctuation state of photovoltaic, wind power and load in the next 24 hours, and assign an equal initial probability of 0.0001 to each scenario. The deviation between the overall probability distribution of the generated scene set and a preset reference distribution fitted by historical data is measured, and its Wasserstein distance is calculated; a confidence radius, such as 0.05, is set, and a fuzzy set is constructed to define all probability distributions whose Wasserstein distance from the reference distribution does not exceed 0.05, thus forming the feasible region of the initial probability distribution. Safety testing was conducted using historical extreme scenarios. Ten extreme daily scenarios that had previously caused system voltage overruns or line overloads were selected from historical data as a set of safety test scenarios. This initial feasible domain was pre-screened, revealing a set of probability distributions. If certain severe weather scenarios were assigned excessively high probabilities, any feasible scheduling strategy would lead to overcharging of energy storage or overruns of critical line power when executing a certain extreme scenario under this distribution. Therefore, this set of risk probability distributions was removed from the initial feasible domain. Through this pre-screening step, the initial fuzzy probability feasible domain was narrowed down to a robust probability distribution feasible domain that ensures scheduling safety, mathematically eliminating the possibility of physical insecurity. Using scene reduction technology, the synchronous back-substitution elimination method was used to reduce 10,000 initial scenes. Similar scenes were merged based on probability distance, and finally reduced to ten key scenes. The adjusted probabilities were recalculated to ensure that their empirical distribution is still within the feasible region of the aforementioned robust probability distribution. A two-stage distributed bar optimization model was established and solved. An optimization objective function was established with the goal of minimizing the total operating cost of the microgrid system. The total cost includes the cost of purchasing and selling electricity, the cost of energy storage loss, the cost of controllable power generation, and the cost of demand response regulation. Specifically, the goal is to minimize the sum of the cost of purchasing electricity, the fuel cost of the micro gas turbine, the cost of energy storage aging, and the cost of demand response compensation. The solution to the objective function must satisfy the system power balance and the physical constraints of the operation of each power generation unit and energy storage device, such as the ramp-up and output limits of the micro gas turbine, the charging and discharging power and energy constraints of energy storage, the state of charge limit, the power exchange limit with the main grid, and the node voltage safety constraint. To address uncertainty, decision variables are divided into two stages. The first stage includes the planned 24-hour output of the micro gas turbine, the planned charging and discharging power of the energy storage, and the day-ahead power purchase plan with the main grid. The second stage addresses each possible uncertainty scenario by adjusting the output of the micro gas turbine, adjusting the charging and discharging of the energy storage, initiating demand response, and adjusting the real-time power exchange with the main grid to balance the deviations between the actual and predicted values ​​of photovoltaic, wind power, and load. A two-stage sub-Bruker optimization model is established, with the goal of finding the optimal scheduling strategy that can cope with the worst probability distribution within the aforementioned robust probability distribution feasible region. The mathematical form of the model is: minimizing the sum of the first-stage cost and the expected second-stage cost under the worst probability distribution within the robust feasible region. Main Problem Solving: Using ten key scenarios and their probabilities as the current estimate of the worst-case distribution, solve a deterministic two-stage optimization problem to obtain candidate solutions for the first-stage decision (such as the gas turbine planning curve) and calculate the lower bound of the objective function. Sub-problem Solving: For the candidate solution, find a new probability distribution within the robust feasible region (either by reallocating scenario probabilities or discovering new worst-case scenarios) such that the expected total cost after executing the predetermined first-stage decision is maximized under this distribution, and calculate this maximum expected cost value as the upper bound. Iteration and Convergence Judgment: Compare the upper and lower bounds. If the difference is less than the preset convergence tolerance (such as one percent), the algorithm terminates, and the current solution is the optimal scheduling strategy; otherwise, add the new worst-case scenario or probability distribution information identified by the sub-problem to the scenario set of the main problem, resolve the main problem, and repeat until convergence. During the iterative solution process, an online evaluation mechanism for the coupling relationship of the two-stage decision variables is introduced simultaneously. When the evaluation shows that strong coupling leads to difficulty in solving the problem, the formulation of the optimization model is dynamically adjusted, and inter-stage decoupling simplification is implemented in some time periods or for some devices to improve the solution efficiency. When the evaluation shows that strong coupling is crucial to ensuring the robustness of the strategy, the coupling relationship is maintained or strengthened. In this embodiment, the evaluation found that the day-ahead planning and real-time adjustment of the energy storage device are closely coupled, so the coupling is maintained. The coupling of the micro gas turbine is weak in some off-peak periods, so decoupling simplification is performed to improve the solution speed of this part. Through multiple iterations until convergence, the candidate optimal scheduling strategy is obtained. Prospective verification and intelligent closed-loop are conducted; before outputting the final scheduling command, simulation is performed to verify the feasibility of the strategy; candidate optimal scheduling strategies (gas turbine plan, energy storage plan, etc.) are input into a high-fidelity simulation environment built based on professional simulation software for prospective simulation; simulation monitoring and recording of all potential problems such as constraint overruns, equipment overloads, voltage and frequency anomalies, as well as corresponding key uncertainty scenario segments and input data characteristics; simulation found that when the execution reaches the 15th hour, if a specific photovoltaic sudden drop and load sudden rise scenario is encountered (the probability of this scenario is underestimated in the ten key scenarios), it will cause the energy storage charging and discharging switching to be too fast, causing the DC bus voltage to momentarily exceed the limit; an anomaly detection report is generated, recording the time, amplitude, and duration of the limit exceedance, and marking the photovoltaic load joint fluctuation mode and corresponding original data characteristics in the 15th hour that caused the problem; The report was used to trace back and locate the source data that caused the strategy problem, including original input outliers, data points that may have been incorrectly corrected in the aforementioned preprocessing, and the generated risk-critical uncertainty scenarios. There were few records of such extreme scenarios with strong correlation between sudden photovoltaic drops and sudden load increases in the historical database, which resulted in the failure to fully characterize their statistical features in the aforementioned probability distribution fitting, and thus they were partially ignored when generating scenarios and safety testing. All problem data and context information were automatically backed up and stored in a dedicated problem data and scenario library, which was physically isolated from the normal data source of the main process. For identified and isolated problematic data, an automatic repair and enhancement process is initiated; for obviously erroneous or missing data, intelligent imputation and correction algorithms are used for repair; for such joint fluctuation pattern records missing in historical data, generative adversarial networks are used for data enhancement to generate reasonable synthetic scenario data to supplement the historical database; for uncertain scenario data that repeatedly causes policy risks but is not obviously erroneous, special marking and enhancement are performed, increasing the penalty weight of the identified risk scenarios in the objective function by 20%, and using adversarial example generation technology to derive several more severe variant scenarios based on them; these enhanced data are used as special stress test scenarios and fed back to the aforementioned feasible domain construction or key scenario set, specifically to the security test scenario set. The closed-loop feedback information generated during the verification, tracing, and repair processes serves as a continuous monitoring signal, driving the adaptive learning of key parameters in the scheduling method. A feedback learning mechanism is established to dynamically adjust core parameters based on the frequency, type, and repair effectiveness of anomalies in historical cycles. Following this anomaly event, the mechanism dynamically adjusts the aforementioned data cleaning threshold, the confidence radius of the feasible region for constructing the probability distribution, the stringency of safety constraints, and the tolerance for model convergence judgment based on recent anomalies caused by insufficient estimation of joint fluctuation patterns. Specifically, this involves appropriately reducing the confidence radius of the feasible region to make it more concentrated, and increasing the voltage safety margin during relevant periods to tighten safety constraints. After performing the aforementioned solution process and comprehensively considering the closed-loop feedback and dynamic adjustment results, the first-stage daytime scheduling plan and the second-stage real-time adjustment strategy, which minimize the total cost of the system under the worst-case probability distribution of uncertainty and pass look-ahead verification, are obtained; this is the optimal scheduling strategy. In the next scheduling cycle, the aforementioned closed-loop feedback has been comprehensively considered when performing model solving. Since safety testing has been added to the stress test scenario of the feedback, the new robust feasible region has excluded such risk distributions. Due to parameter adjustments, the model is more conservative. After resolving, a new scheduling plan and adjustment strategy are obtained, and simulation verification shows that there is no voltage limit exceeding problem; this is the final optimal scheduling strategy. Output dispatch instructions for each controllable power generation unit, energy storage system and demand response resources; decompose the final strategy into specific instructions such as the 24-hour start-up and shutdown and output curve of the micro gas turbine, the 24-hour charging and discharging plan of the energy storage system, the contracted electricity with the main grid, and the preparatory plan for the call of demand response resources, and issue them to each execution unit.

[0038] The specific data are shown in Tables 1 and 2 below. These tables show the operational details of the method itself, including the effect of prediction error correction and solution efficiency.

[0039] Table 1: Excerpt of operational data of the method of the present invention over 30 consecutive test days;

[0040] Table 2: Comparison of the overall performance of the method of the present invention and the existing technology method within 30 test days;

[0041] Based on the comparative data in Tables 1 and 2 above, the superiority of the method proposed in this invention is mainly reflected in the following three aspects: 1. The non-dominated balance between safety and economy, i.e., Pareto optimality: As shown in Table 2, while traditional stochastic optimization (SO) appears to be the most economical, its voltage over-limit probability is as high as 18.5%, which frequently triggers protection devices in actual operation, leading to serious safety accidents. Its low cost comes at the expense of reliability. Traditional robust optimization (RO), while guaranteeing zero over-limits, is extremely costly and economically inefficient. The method of this invention, through the robust probability distribution feasible region contraction technique in step S3, eliminates extreme distributions that inevitably lead to over-limits, enabling the optimization result to maintain economic efficiency close to DRO while achieving the same zero over-limit safety standard as RO. The cost standard deviation is only 985.4, indicating that the expenditure fluctuation of this method is minimal under different operating conditions, demonstrating excellent robustness.

[0042] 2. Data-driven closed-loop adaptive evolution capability: Standard DRO still has a 5.2% failure rate when encountering unmodeled extreme scenarios such as the thunderstorms on Day 13, classifying it as a "static model." The unique S7-S10 closed-loop mechanism of this invention, after the early warning on Day 13, traces the source of the problem in S8, repairs and enhances the data in S9, and automatically adjusts the confidence radius parameter for subsequent tests in S10. This explains why the method of this invention can still maintain zero failure rates under similar weather conditions in the future. This self-learning characteristic of "getting smarter with each run" is a core advantage that other static optimization algorithms do not possess.

[0043] 3. Intelligent balance between solution efficiency and accuracy: Although the model complexity of this invention is higher than SO and RO, the solution time is significantly better than the standard DR due to the introduction of an online evaluation and dynamic decoupling mechanism for the coupling relationship in steps S64 and S65, and is far below the engineering-allowed 15-minute scheduling cycle. While ensuring the convergence of the global optimal solution, it achieves efficient utilization of computing resources and possesses extremely high engineering practical value.

[0044] like Figure 3 As shown, a violin plot-style visualization technique, combining box plots with scatter plots, is used to visually demonstrate the distribution of maximum node voltages during 30 independent random operating days of testing, comparing the method of this invention with three existing technologies. The red dashed line in the figure represents the upper limit threshold for safe voltage operation.

[0045] As clearly seen in the figure, the data points of traditional stochastic optimization (SO) are extremely dispersed, with a large number of sample points located above the red dashed line, and the upper edge of the box far exceeding the safety limit. This indicates that although the method is computationally simple, it lacks robustness in the face of fluctuations in renewable energy output, posing a very high risk of voltage exceeding limits. The data points of traditional robust optimization (RO) are extremely densely clustered below the safety line, verifying its extreme conservatism of "designing for the worst-case scenario." While absolutely safe, this comes at a huge economic cost. The distribution of standard distributed robust optimization (DRO) falls between the two, but some outliers still touch or slightly exceed the safety line, indicating that its probabilistic fuzzy set fails to fully cover actual physical risks.

[0046] In contrast, the data distribution of the method of this invention exhibits three significant characteristics: first, the median is extremely low and close to the ideal value of 1.0; second, the box span is extremely small, indicating excellent strategy stability and immunity to random disturbances; and third, all scatter points are strictly within the safety red line, with no points exceeding the limit. This superior performance is attributed to the safety-compacting robust feasible region construction technique in step S3 of claim 1, which eliminates probability distributions that could lead to exceeding the limit from the mathematical source, enabling the optimization results to achieve a safety guarantee comparable to or even surpassing traditional robust optimization while preserving the economy of partial robustness. The figure demonstrates from a statistical empirical perspective that the method possesses first-class robustness and control accuracy under uncertain environments.

[0047] Figure 4 Using grouped bar charts, a normalized quantitative score comparison was conducted on the four core dimensions of greatest concern in microgrid economic dispatch: operational economy, voltage safety, solution efficiency, and renewable energy integration rate. A score closer to 1.0 indicates a better overall performance of the method on that indicator.

[0048] like Figure 4 As shown, traditional stochastic optimization (SO) scores highly in operational economy but scores almost zero in voltage safety, exhibiting a serious performance weakness. Traditional robust optimization (RO), on the other hand, scores full marks in safety but has extremely low economic efficiency, demonstrating a typical imbalance. While standard-scoring robust optimization (DRO) is above average in all three metrics, its solution efficiency score is the lowest, reflecting the complexity of its model solution.

[0049] In contrast, the method of this invention consistently scores above 0.7 in all four dimensions. Specifically: through the feasible region contraction technique in step S3, this method achieves full security with minimal economic cost; due to the introduction of the dynamic decoupling evaluation mechanism in steps S64-S65, its computational time score is significantly better than the standard DRO, greatly improving engineering usability while ensuring solution quality; and the radar chart area of ​​this method is significantly larger than the other three methods in all four dimensions.

[0050] Example 3: This embodiment provides a system for implementing the above method. This system can be integrated into a microgrid central controller (MGCC) or a cloud platform, including: The data acquisition and preprocessing module is used to execute steps S1 / A1-A4; The uncertainty scenario management and feasible domain construction module is used to execute steps S3 and S4 and contains a safety testing engine. Two-stage decomposer optimization solver: used to execute steps S2, S5, and S6, containing an iterative decomposition algorithm and a coupled evaluator; The high-fidelity forward-looking simulation verification module is used to execute step S7 and interfaces with simulation software such as MATLAB / Simulink and RT-LAB. The data tracing and intelligent repair module is used to execute steps S8 and S9; The parameter adaptive learning engine is used to execute step S10 to achieve dynamic parameter adjustment; The scheduling instruction management and output module is used to execute steps S11 and S12 and communicate with each execution unit. The modules are connected in sequence according to the process flow, and work together to form a complete intelligent scheduling system.

Claims

1. A robust and economical dispatch method for microgrids based on probability distribution sets, characterized in that, include: S1 acquires and standardizes historical and real-time photovoltaic and wind power output and load data, and quantifies uncertainty characteristics by analyzing the data fluctuation patterns and statistical properties. S2, establish an optimization objective function with the goal of minimizing the total operating cost of the microgrid system; S3. Construct a feasible region of a safe and robust probability distribution for subsequent robust optimization. First, generate uncertain scenarios and construct an initial fuzzy set describing the possible fluctuation range of the probability distribution. Through safety testing based on extreme scenarios, remove all risk probability distributions that will inevitably lead to the system running out of limits from this set, thus forming a robust probability distribution feasible region that eliminates safety risks from the source and is compacted. S4. After generating a large number of initial uncertain scenarios through step S3, the numerous initial scenarios are merged and filtered according to their probability characteristics through scenario reduction technology, and finally reduced to a limited number of representative key scenarios, and their corresponding adjusted probabilities are calculated. S5. Establish a two-stage sub-Bruker optimization model and solve the optimal scheduling strategy for dealing with the worst probability distribution of uncertainty within the feasible region of the robust probability distribution. S6. The iterative decomposition algorithm is used to solve the two-stage sub-Brussels bar optimization model established in step S5. S7 performs simulation processing before outputting the final scheduling instruction to verify the feasibility of the strategy and generate an anomaly detection report. S8. Using the anomaly detection report generated in step S7, trace back and locate the source data that caused the scheduling strategy problem. S9. For the problematic data identified and isolated in step S8, initiate the automatic repair and enhancement process. S10, the closed-loop feedback information formed in steps S7 to S9 is used as a continuous supervision signal to drive the adaptive learning of key parameters in the scheduling method. S11, execute the solution process of step S6, and after comprehensively considering the closed-loop feedback and dynamic adjustment results of steps S7 to S10, obtain the first-stage day-ahead scheduling plan and the second-stage real-time adjustment strategy that minimize the total cost of the system under the worst probability distribution of uncertainty and pass the look-ahead verification, i.e. the optimal scheduling strategy. S12, Output dispatch instructions for each controllable power generation unit, energy storage system and demand response resource.

2. The robust economic dispatch method for microgrids based on probability distribution sets as described in claim 1, characterized in that, Step S1 specifically includes the following steps: A1. Obtain historical and real-time photovoltaic power output data, wind power output data, and load demand data to form the raw input dataset; A2. Perform preprocessing on the original input dataset. Preprocessing specifically includes: A21. Identify and remove obvious errors and outliers from the data; A22. For missing time periods in the data, use interpolation of nearby data or filling methods based on the historical average of similar days to complete the data; A23. Scale all data to the [0,1] interval according to their respective maximum and minimum value ranges to form a standardized input data sequence; A3. Calculate the statistical characteristics of uncertainty parameters using standardized input data sequences: A31. Calculate the historical maximum and minimum value ranges of photovoltaic power output, wind power output, and load demand during typical hourly periods; A32. Calculate the Pearson correlation coefficient matrix between photovoltaic and wind power output, and between renewable energy output and load demand; A33. For the prediction errors and load fluctuations of photovoltaic and wind power respectively, the probability distributions are fitted using historical data. The distributions include, but are not limited to, normal distribution, Beta distribution, and Gaussian mixture distribution, and the probability characteristics are characterized by the distribution parameters. A4. Directly output the processed standardized data sequence, as well as the set of uncertain statistical features of photovoltaic, wind power output and load demand, including their fluctuation range, correlation coefficient matrix and fitted probability distribution parameters, as input for generating uncertain scenarios.

3. The robust economic dispatch method for microgrids based on probability distribution sets as described in claim 1, characterized in that, The total operating cost in step S2 specifically includes the cost of purchasing and selling electricity, the cost of energy storage equipment losses, the generation cost of controllable distributed power sources, and the regulation cost generated by implementing demand response. The solution of the objective function needs to satisfy the system power balance constraints, the physical constraints of operation of each power generation unit and energy storage equipment, and the safety constraints to ensure the operation of the system.

4. The robust economic dispatch method for microgrids based on probability distribution sets as described in claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Using the probability distribution features extracted in step S1, generate typical scenarios that characterize the possible future states of the uncertainty parameters, and assign an initial probability estimate to each scenario. S32, measure the deviation between the overall probability distribution of the generated scene set and a preset reference distribution, and construct a fuzzy set that allows the true probability to fluctuate around the reference distribution. The fuzzy set defines the feasible region of the initial probability distribution formed by all possible probability distributions. S33, using historical extreme scenarios or a pre-set set of security test scenarios, pre-screen the feasible region of the initial probability distribution, and remove those probability distributions from the feasible region that, although they meet the definition of fuzzy sets, would cause any feasible scheduling strategy to violate key security constraints; S34, through a pre-screening step, compresses the initial fuzzy probability feasible region into a robust probability distribution feasible region that ensures scheduling safety, providing an uncertainty description framework.

5. The robust economic dispatch method for microgrids based on probability distribution sets as described in claim 1, characterized in that, Step S5 specifically includes the following steps: S51, in order to make decisions in response to uncertainty, the scheduling decision variables are divided into two stages; S52, the first stage is the day-ahead decision variable, which is determined before the uncertainty is revealed; the second stage is the real-time adjustment variable, which is adjusted after the uncertainty is realized, and a two-stage sub-Bruker optimization model is established. S53, the objective of the sub-Bruker optimization model is to find the optimal scheduling strategy that can cope with the worst probability distribution situation within the feasible region of the robust probability distribution defined in step S3. S54, the mathematical form of the sub-Bruker optimization model is as follows: determine the first-stage decision variables and their costs; for the first-stage decision and any possible uncertainty scenario, optimize the second-stage decision variables to minimize the operating cost under that scenario; the optimization objective is to minimize the sum of the first-stage cost and the expected second-stage cost under the worst probability distribution in the feasible region of the robust probability distribution.

6. The robust economic dispatch method for microgrids based on probability distribution sets as described in claim 1, characterized in that, Step S6 specifically includes the following steps: S61, Solving the main problem: Using the key scenario and adjusted probability of the current step S4, solve the deterministic two-stage optimization problem, obtain the candidate solution for the first-stage decision, and calculate the lower bound of the objective function; S62, Subproblem solving: For the first-stage candidate solution given by the main problem, within the feasible region of the robust probability distribution defined in step S3, find the worst probability distribution that maximizes the total expected cost, and calculate the corresponding expected cost value to form the upper bound of the objective function; S63, Iteration and Convergence: Compare the currently calculated upper bound with the lower bound. If the difference is less than the preset convergence tolerance, the algorithm terminates and the current solution is the optimal scheduling strategy. Otherwise, add the new worst-case scenario or the corresponding probability distribution information identified by the subproblem to the scenario set of the main problem, solve the main problem again, and repeat the above process until convergence to seek the global optimal solution. S64 introduces an online evaluation mechanism for the coupling relationship between the day-ahead decision in the first stage and the real-time adjustment decision variables in the second stage during the iterative solution process. S65. When the evaluation shows that strong coupling between the two leads to difficulty in solving the problem, the expression of the optimization model is dynamically adjusted, and a decoupling and simplification strategy is implemented between stages in some time periods or for some devices to improve the solution efficiency. When the evaluation shows that strong coupling is crucial to ensuring the robustness of the strategy, the coupling relationship is maintained or strengthened.

7. The robust economic dispatch method for microgrids based on probability distribution sets as described in claim 1, characterized in that, Step S7 specifically includes the following steps: S71, Before outputting the final scheduling instruction in step S6, the generated candidate optimal scheduling strategy is prospectively simulated in a high-fidelity simulation environment built based on the current data; S72, the forward-looking simulation simultaneously monitors and records all potential problems that may occur during the strategy execution process, such as constraint overruns, equipment overloads, and voltage / frequency anomalies, as well as key uncertainty scenario segments and input data characteristics that lead to these problems; S73, finally generates an anomaly detection report.

8. A robust and economical microgrid dispatch method based on probability distribution sets as described in claim 7, characterized in that, The source data includes outliers in the original input data, data points that may be incorrectly corrected in the preprocessing step S1, and key uncertainty scenarios that lead to risks. All identified problem data and context information are automatically backed up and stored in the problem data and scenario library, which is physically or logically isolated from the normal data sources used in the main process.

9. A robust and economical dispatch method for microgrids based on probability distribution sets as described in claim 1, characterized in that, Step S9 specifically includes the following steps: S91 initiates an automatic repair and enhancement process for problematic data. For data with obvious errors or missing information, it uses intelligent interpolation and correction algorithms for repair. S92. For uncertain scenario data that repeatedly causes policy risks but is not obvious errors, special labeling and enhancement are performed by increasing the corresponding penalty weight in the objective function or generating adversarial samples based on it. S93, these enhanced data are used as a special stress test scenario and fed back to the probability distribution feasible domain construction in step S3 or the key scenario set in step S4.

10. A robust and economical microgrid dispatch method based on probability distribution sets as described in claim 1, characterized in that, Step S10 specifically includes the following steps: S101. The closed-loop feedback information formed in steps S7 to S9 is used as a continuous monitoring signal to drive the adaptive learning of key parameters in the scheduling method. S102. Establish a feedback learning mechanism to dynamically adjust the data cleaning threshold in step S1, the confidence radius of the feasible region of the probability distribution in step S3, the strictness of the safety constraints in step S3, and the tolerance core parameters of the model convergence judgment in step S6, based on the frequency, type, and repair effect of anomalies in the historical cycle. S103, enabling the entire scheduling system to continuously learn from historical problems and optimize itself.

Images